1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2018 Intel Corporation. */
3 
4 #include <linux/bpf_trace.h>
5 #include <net/xdp_sock_drv.h>
6 #include <net/xdp.h>
7 
8 #include "ixgbe.h"
9 #include "ixgbe_txrx_common.h"
10 
11 struct xsk_buff_pool *ixgbe_xsk_pool(struct ixgbe_adapter *adapter,
12 				     struct ixgbe_ring *ring)
13 {
14 	bool xdp_on = READ_ONCE(adapter->xdp_prog);
15 	int qid = ring->ring_idx;
16 
17 	if (!xdp_on || !test_bit(qid, adapter->af_xdp_zc_qps))
18 		return NULL;
19 
20 	return xsk_get_pool_from_qid(adapter->netdev, qid);
21 }
22 
23 static int ixgbe_xsk_pool_enable(struct ixgbe_adapter *adapter,
24 				 struct xsk_buff_pool *pool,
25 				 u16 qid)
26 {
27 	struct net_device *netdev = adapter->netdev;
28 	bool if_running;
29 	int err;
30 
31 	if (qid >= adapter->num_rx_queues)
32 		return -EINVAL;
33 
34 	if (qid >= netdev->real_num_rx_queues ||
35 	    qid >= netdev->real_num_tx_queues)
36 		return -EINVAL;
37 
38 	err = xsk_pool_dma_map(pool, &adapter->pdev->dev, IXGBE_RX_DMA_ATTR);
39 	if (err)
40 		return err;
41 
42 	if_running = netif_running(adapter->netdev) &&
43 		     ixgbe_enabled_xdp_adapter(adapter);
44 
45 	if (if_running)
46 		ixgbe_txrx_ring_disable(adapter, qid);
47 
48 	set_bit(qid, adapter->af_xdp_zc_qps);
49 
50 	if (if_running) {
51 		ixgbe_txrx_ring_enable(adapter, qid);
52 
53 		/* Kick start the NAPI context so that receiving will start */
54 		err = ixgbe_xsk_wakeup(adapter->netdev, qid, XDP_WAKEUP_RX);
55 		if (err)
56 			return err;
57 	}
58 
59 	return 0;
60 }
61 
62 static int ixgbe_xsk_pool_disable(struct ixgbe_adapter *adapter, u16 qid)
63 {
64 	struct xsk_buff_pool *pool;
65 	bool if_running;
66 
67 	pool = xsk_get_pool_from_qid(adapter->netdev, qid);
68 	if (!pool)
69 		return -EINVAL;
70 
71 	if_running = netif_running(adapter->netdev) &&
72 		     ixgbe_enabled_xdp_adapter(adapter);
73 
74 	if (if_running)
75 		ixgbe_txrx_ring_disable(adapter, qid);
76 
77 	clear_bit(qid, adapter->af_xdp_zc_qps);
78 	xsk_pool_dma_unmap(pool, IXGBE_RX_DMA_ATTR);
79 
80 	if (if_running)
81 		ixgbe_txrx_ring_enable(adapter, qid);
82 
83 	return 0;
84 }
85 
86 int ixgbe_xsk_pool_setup(struct ixgbe_adapter *adapter,
87 			 struct xsk_buff_pool *pool,
88 			 u16 qid)
89 {
90 	return pool ? ixgbe_xsk_pool_enable(adapter, pool, qid) :
91 		ixgbe_xsk_pool_disable(adapter, qid);
92 }
93 
94 static int ixgbe_run_xdp_zc(struct ixgbe_adapter *adapter,
95 			    struct ixgbe_ring *rx_ring,
96 			    struct xdp_buff *xdp)
97 {
98 	int err, result = IXGBE_XDP_PASS;
99 	struct bpf_prog *xdp_prog;
100 	struct xdp_frame *xdpf;
101 	u32 act;
102 
103 	rcu_read_lock();
104 	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
105 	act = bpf_prog_run_xdp(xdp_prog, xdp);
106 
107 	switch (act) {
108 	case XDP_PASS:
109 		break;
110 	case XDP_TX:
111 		xdpf = xdp_convert_buff_to_frame(xdp);
112 		if (unlikely(!xdpf)) {
113 			result = IXGBE_XDP_CONSUMED;
114 			break;
115 		}
116 		result = ixgbe_xmit_xdp_ring(adapter, xdpf);
117 		break;
118 	case XDP_REDIRECT:
119 		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
120 		result = !err ? IXGBE_XDP_REDIR : IXGBE_XDP_CONSUMED;
121 		break;
122 	default:
123 		bpf_warn_invalid_xdp_action(act);
124 		fallthrough;
125 	case XDP_ABORTED:
126 		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
127 		fallthrough; /* handle aborts by dropping packet */
128 	case XDP_DROP:
129 		result = IXGBE_XDP_CONSUMED;
130 		break;
131 	}
132 	rcu_read_unlock();
133 	return result;
134 }
135 
136 bool ixgbe_alloc_rx_buffers_zc(struct ixgbe_ring *rx_ring, u16 count)
137 {
138 	union ixgbe_adv_rx_desc *rx_desc;
139 	struct ixgbe_rx_buffer *bi;
140 	u16 i = rx_ring->next_to_use;
141 	dma_addr_t dma;
142 	bool ok = true;
143 
144 	/* nothing to do */
145 	if (!count)
146 		return true;
147 
148 	rx_desc = IXGBE_RX_DESC(rx_ring, i);
149 	bi = &rx_ring->rx_buffer_info[i];
150 	i -= rx_ring->count;
151 
152 	do {
153 		bi->xdp = xsk_buff_alloc(rx_ring->xsk_pool);
154 		if (!bi->xdp) {
155 			ok = false;
156 			break;
157 		}
158 
159 		dma = xsk_buff_xdp_get_dma(bi->xdp);
160 
161 		/* Refresh the desc even if buffer_addrs didn't change
162 		 * because each write-back erases this info.
163 		 */
164 		rx_desc->read.pkt_addr = cpu_to_le64(dma);
165 
166 		rx_desc++;
167 		bi++;
168 		i++;
169 		if (unlikely(!i)) {
170 			rx_desc = IXGBE_RX_DESC(rx_ring, 0);
171 			bi = rx_ring->rx_buffer_info;
172 			i -= rx_ring->count;
173 		}
174 
175 		/* clear the length for the next_to_use descriptor */
176 		rx_desc->wb.upper.length = 0;
177 
178 		count--;
179 	} while (count);
180 
181 	i += rx_ring->count;
182 
183 	if (rx_ring->next_to_use != i) {
184 		rx_ring->next_to_use = i;
185 
186 		/* Force memory writes to complete before letting h/w
187 		 * know there are new descriptors to fetch.  (Only
188 		 * applicable for weak-ordered memory model archs,
189 		 * such as IA-64).
190 		 */
191 		wmb();
192 		writel(i, rx_ring->tail);
193 	}
194 
195 	return ok;
196 }
197 
198 static struct sk_buff *ixgbe_construct_skb_zc(struct ixgbe_ring *rx_ring,
199 					      struct ixgbe_rx_buffer *bi)
200 {
201 	unsigned int metasize = bi->xdp->data - bi->xdp->data_meta;
202 	unsigned int datasize = bi->xdp->data_end - bi->xdp->data;
203 	struct sk_buff *skb;
204 
205 	/* allocate a skb to store the frags */
206 	skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
207 			       bi->xdp->data_end - bi->xdp->data_hard_start,
208 			       GFP_ATOMIC | __GFP_NOWARN);
209 	if (unlikely(!skb))
210 		return NULL;
211 
212 	skb_reserve(skb, bi->xdp->data - bi->xdp->data_hard_start);
213 	memcpy(__skb_put(skb, datasize), bi->xdp->data, datasize);
214 	if (metasize)
215 		skb_metadata_set(skb, metasize);
216 
217 	xsk_buff_free(bi->xdp);
218 	bi->xdp = NULL;
219 	return skb;
220 }
221 
222 static void ixgbe_inc_ntc(struct ixgbe_ring *rx_ring)
223 {
224 	u32 ntc = rx_ring->next_to_clean + 1;
225 
226 	ntc = (ntc < rx_ring->count) ? ntc : 0;
227 	rx_ring->next_to_clean = ntc;
228 	prefetch(IXGBE_RX_DESC(rx_ring, ntc));
229 }
230 
231 int ixgbe_clean_rx_irq_zc(struct ixgbe_q_vector *q_vector,
232 			  struct ixgbe_ring *rx_ring,
233 			  const int budget)
234 {
235 	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
236 	struct ixgbe_adapter *adapter = q_vector->adapter;
237 	u16 cleaned_count = ixgbe_desc_unused(rx_ring);
238 	unsigned int xdp_res, xdp_xmit = 0;
239 	bool failure = false;
240 	struct sk_buff *skb;
241 
242 	while (likely(total_rx_packets < budget)) {
243 		union ixgbe_adv_rx_desc *rx_desc;
244 		struct ixgbe_rx_buffer *bi;
245 		unsigned int size;
246 
247 		/* return some buffers to hardware, one at a time is too slow */
248 		if (cleaned_count >= IXGBE_RX_BUFFER_WRITE) {
249 			failure = failure ||
250 				  !ixgbe_alloc_rx_buffers_zc(rx_ring,
251 							     cleaned_count);
252 			cleaned_count = 0;
253 		}
254 
255 		rx_desc = IXGBE_RX_DESC(rx_ring, rx_ring->next_to_clean);
256 		size = le16_to_cpu(rx_desc->wb.upper.length);
257 		if (!size)
258 			break;
259 
260 		/* This memory barrier is needed to keep us from reading
261 		 * any other fields out of the rx_desc until we know the
262 		 * descriptor has been written back
263 		 */
264 		dma_rmb();
265 
266 		bi = &rx_ring->rx_buffer_info[rx_ring->next_to_clean];
267 
268 		if (unlikely(!ixgbe_test_staterr(rx_desc,
269 						 IXGBE_RXD_STAT_EOP))) {
270 			struct ixgbe_rx_buffer *next_bi;
271 
272 			xsk_buff_free(bi->xdp);
273 			bi->xdp = NULL;
274 			ixgbe_inc_ntc(rx_ring);
275 			next_bi =
276 			       &rx_ring->rx_buffer_info[rx_ring->next_to_clean];
277 			next_bi->discard = true;
278 			continue;
279 		}
280 
281 		if (unlikely(bi->discard)) {
282 			xsk_buff_free(bi->xdp);
283 			bi->xdp = NULL;
284 			bi->discard = false;
285 			ixgbe_inc_ntc(rx_ring);
286 			continue;
287 		}
288 
289 		bi->xdp->data_end = bi->xdp->data + size;
290 		xsk_buff_dma_sync_for_cpu(bi->xdp, rx_ring->xsk_pool);
291 		xdp_res = ixgbe_run_xdp_zc(adapter, rx_ring, bi->xdp);
292 
293 		if (xdp_res) {
294 			if (xdp_res & (IXGBE_XDP_TX | IXGBE_XDP_REDIR))
295 				xdp_xmit |= xdp_res;
296 			else
297 				xsk_buff_free(bi->xdp);
298 
299 			bi->xdp = NULL;
300 			total_rx_packets++;
301 			total_rx_bytes += size;
302 
303 			cleaned_count++;
304 			ixgbe_inc_ntc(rx_ring);
305 			continue;
306 		}
307 
308 		/* XDP_PASS path */
309 		skb = ixgbe_construct_skb_zc(rx_ring, bi);
310 		if (!skb) {
311 			rx_ring->rx_stats.alloc_rx_buff_failed++;
312 			break;
313 		}
314 
315 		cleaned_count++;
316 		ixgbe_inc_ntc(rx_ring);
317 
318 		if (eth_skb_pad(skb))
319 			continue;
320 
321 		total_rx_bytes += skb->len;
322 		total_rx_packets++;
323 
324 		ixgbe_process_skb_fields(rx_ring, rx_desc, skb);
325 		ixgbe_rx_skb(q_vector, skb);
326 	}
327 
328 	if (xdp_xmit & IXGBE_XDP_REDIR)
329 		xdp_do_flush_map();
330 
331 	if (xdp_xmit & IXGBE_XDP_TX) {
332 		struct ixgbe_ring *ring = adapter->xdp_ring[smp_processor_id()];
333 
334 		/* Force memory writes to complete before letting h/w
335 		 * know there are new descriptors to fetch.
336 		 */
337 		wmb();
338 		writel(ring->next_to_use, ring->tail);
339 	}
340 
341 	u64_stats_update_begin(&rx_ring->syncp);
342 	rx_ring->stats.packets += total_rx_packets;
343 	rx_ring->stats.bytes += total_rx_bytes;
344 	u64_stats_update_end(&rx_ring->syncp);
345 	q_vector->rx.total_packets += total_rx_packets;
346 	q_vector->rx.total_bytes += total_rx_bytes;
347 
348 	if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
349 		if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
350 			xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
351 		else
352 			xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
353 
354 		return (int)total_rx_packets;
355 	}
356 	return failure ? budget : (int)total_rx_packets;
357 }
358 
359 void ixgbe_xsk_clean_rx_ring(struct ixgbe_ring *rx_ring)
360 {
361 	struct ixgbe_rx_buffer *bi;
362 	u16 i;
363 
364 	for (i = 0; i < rx_ring->count; i++) {
365 		bi = &rx_ring->rx_buffer_info[i];
366 
367 		if (!bi->xdp)
368 			continue;
369 
370 		xsk_buff_free(bi->xdp);
371 		bi->xdp = NULL;
372 	}
373 }
374 
375 static bool ixgbe_xmit_zc(struct ixgbe_ring *xdp_ring, unsigned int budget)
376 {
377 	struct xsk_buff_pool *pool = xdp_ring->xsk_pool;
378 	union ixgbe_adv_tx_desc *tx_desc = NULL;
379 	struct ixgbe_tx_buffer *tx_bi;
380 	bool work_done = true;
381 	struct xdp_desc desc;
382 	dma_addr_t dma;
383 	u32 cmd_type;
384 
385 	while (budget-- > 0) {
386 		if (unlikely(!ixgbe_desc_unused(xdp_ring)) ||
387 		    !netif_carrier_ok(xdp_ring->netdev)) {
388 			work_done = false;
389 			break;
390 		}
391 
392 		if (!xsk_tx_peek_desc(pool, &desc))
393 			break;
394 
395 		dma = xsk_buff_raw_get_dma(pool, desc.addr);
396 		xsk_buff_raw_dma_sync_for_device(pool, dma, desc.len);
397 
398 		tx_bi = &xdp_ring->tx_buffer_info[xdp_ring->next_to_use];
399 		tx_bi->bytecount = desc.len;
400 		tx_bi->xdpf = NULL;
401 		tx_bi->gso_segs = 1;
402 
403 		tx_desc = IXGBE_TX_DESC(xdp_ring, xdp_ring->next_to_use);
404 		tx_desc->read.buffer_addr = cpu_to_le64(dma);
405 
406 		/* put descriptor type bits */
407 		cmd_type = IXGBE_ADVTXD_DTYP_DATA |
408 			   IXGBE_ADVTXD_DCMD_DEXT |
409 			   IXGBE_ADVTXD_DCMD_IFCS;
410 		cmd_type |= desc.len | IXGBE_TXD_CMD;
411 		tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
412 		tx_desc->read.olinfo_status =
413 			cpu_to_le32(desc.len << IXGBE_ADVTXD_PAYLEN_SHIFT);
414 
415 		xdp_ring->next_to_use++;
416 		if (xdp_ring->next_to_use == xdp_ring->count)
417 			xdp_ring->next_to_use = 0;
418 	}
419 
420 	if (tx_desc) {
421 		ixgbe_xdp_ring_update_tail(xdp_ring);
422 		xsk_tx_release(pool);
423 	}
424 
425 	return !!budget && work_done;
426 }
427 
428 static void ixgbe_clean_xdp_tx_buffer(struct ixgbe_ring *tx_ring,
429 				      struct ixgbe_tx_buffer *tx_bi)
430 {
431 	xdp_return_frame(tx_bi->xdpf);
432 	dma_unmap_single(tx_ring->dev,
433 			 dma_unmap_addr(tx_bi, dma),
434 			 dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
435 	dma_unmap_len_set(tx_bi, len, 0);
436 }
437 
438 bool ixgbe_clean_xdp_tx_irq(struct ixgbe_q_vector *q_vector,
439 			    struct ixgbe_ring *tx_ring, int napi_budget)
440 {
441 	u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
442 	unsigned int total_packets = 0, total_bytes = 0;
443 	struct xsk_buff_pool *pool = tx_ring->xsk_pool;
444 	union ixgbe_adv_tx_desc *tx_desc;
445 	struct ixgbe_tx_buffer *tx_bi;
446 	u32 xsk_frames = 0;
447 
448 	tx_bi = &tx_ring->tx_buffer_info[ntc];
449 	tx_desc = IXGBE_TX_DESC(tx_ring, ntc);
450 
451 	while (ntc != ntu) {
452 		if (!(tx_desc->wb.status & cpu_to_le32(IXGBE_TXD_STAT_DD)))
453 			break;
454 
455 		total_bytes += tx_bi->bytecount;
456 		total_packets += tx_bi->gso_segs;
457 
458 		if (tx_bi->xdpf)
459 			ixgbe_clean_xdp_tx_buffer(tx_ring, tx_bi);
460 		else
461 			xsk_frames++;
462 
463 		tx_bi->xdpf = NULL;
464 
465 		tx_bi++;
466 		tx_desc++;
467 		ntc++;
468 		if (unlikely(ntc == tx_ring->count)) {
469 			ntc = 0;
470 			tx_bi = tx_ring->tx_buffer_info;
471 			tx_desc = IXGBE_TX_DESC(tx_ring, 0);
472 		}
473 
474 		/* issue prefetch for next Tx descriptor */
475 		prefetch(tx_desc);
476 	}
477 
478 	tx_ring->next_to_clean = ntc;
479 
480 	u64_stats_update_begin(&tx_ring->syncp);
481 	tx_ring->stats.bytes += total_bytes;
482 	tx_ring->stats.packets += total_packets;
483 	u64_stats_update_end(&tx_ring->syncp);
484 	q_vector->tx.total_bytes += total_bytes;
485 	q_vector->tx.total_packets += total_packets;
486 
487 	if (xsk_frames)
488 		xsk_tx_completed(pool, xsk_frames);
489 
490 	if (xsk_uses_need_wakeup(pool))
491 		xsk_set_tx_need_wakeup(pool);
492 
493 	return ixgbe_xmit_zc(tx_ring, q_vector->tx.work_limit);
494 }
495 
496 int ixgbe_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags)
497 {
498 	struct ixgbe_adapter *adapter = netdev_priv(dev);
499 	struct ixgbe_ring *ring;
500 
501 	if (test_bit(__IXGBE_DOWN, &adapter->state))
502 		return -ENETDOWN;
503 
504 	if (!READ_ONCE(adapter->xdp_prog))
505 		return -ENXIO;
506 
507 	if (qid >= adapter->num_xdp_queues)
508 		return -ENXIO;
509 
510 	ring = adapter->xdp_ring[qid];
511 
512 	if (test_bit(__IXGBE_TX_DISABLED, &ring->state))
513 		return -ENETDOWN;
514 
515 	if (!ring->xsk_pool)
516 		return -ENXIO;
517 
518 	if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi)) {
519 		u64 eics = BIT_ULL(ring->q_vector->v_idx);
520 
521 		ixgbe_irq_rearm_queues(adapter, eics);
522 	}
523 
524 	return 0;
525 }
526 
527 void ixgbe_xsk_clean_tx_ring(struct ixgbe_ring *tx_ring)
528 {
529 	u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
530 	struct xsk_buff_pool *pool = tx_ring->xsk_pool;
531 	struct ixgbe_tx_buffer *tx_bi;
532 	u32 xsk_frames = 0;
533 
534 	while (ntc != ntu) {
535 		tx_bi = &tx_ring->tx_buffer_info[ntc];
536 
537 		if (tx_bi->xdpf)
538 			ixgbe_clean_xdp_tx_buffer(tx_ring, tx_bi);
539 		else
540 			xsk_frames++;
541 
542 		tx_bi->xdpf = NULL;
543 
544 		ntc++;
545 		if (ntc == tx_ring->count)
546 			ntc = 0;
547 	}
548 
549 	if (xsk_frames)
550 		xsk_tx_completed(pool, xsk_frames);
551 }
552